待拟合曲线为,noise为引入的噪声:
现在打算用一个多项式来拟合这个函数:
损失计算采用平方损失函数:
各个权重的偏导容易计算得出为:
代码如下:
import numpy as np
import matplotlib.pyplot as plt
def init_data():
# 待拟合曲线f(x)=cos2x*[(x^2-1)^3+0.5]
xdata = np.arange(-1, 1, 0.01)
y = [((x ** 2 - 1) ** 3 + 0.5) * np.cos(2 * x) for x in xdata]
ydata = []
for i in range(len(y)):
# 噪声(-0.1,0.1)
noise = np.random.randint(low=-10, high=10) / 100
ydata.append(y[i] + noise)
return xdata, ydata
def loss(x, y, w):
error = 0.0
# 这里使用的是平方损失函数,对应上述公式还是很好理解的
# 这里对每个x的误差都累计起来
for i in range(len(x)):
yl = 0.0
for k in range(len(w)):
yl += w[k] * x[i] ** k
error += (yl - y[i]) ** 2
return error
def gradient_descent(xdata, ydata, pow, epoch=1000, learning_rate=0.001):
'''
:param xdata: x数据
:param ydata: y数据
:param pow: 拟合多项式的幂
:param epoch: 迭代次数
:param learning_rate: 学习率
:return: 权重系数
'''
# 权重初始化为0
w = [0.0] * pow
for r in range(epoch):
for k in range(len(w)):
gradient = 0.0
for i in range(len(xdata)):
yl = 0.0
for j in range(len(w)):
yl += w[j] * x[i] ** j
# 当前权重的平方损失的梯度
gradient += -2 * (ydata[i] - yl) * xdata[i] ** k
w[k] = w[k] - gradient * learning_rate
if r % 100 == 0:
error = loss(xdata, ydata, w)
print("epoch:{}, loss:{}".format(r, error))
return w
x, y = init_data()
w = gradient_descent(x, y, 10)
xpredict = np.arange(-1, 1, 0.01)
ypredict = []
for i in range(len(xpredict)):
yl = 0.0
for j in range(len(w)):
yl += w[j] * xpredict[i] ** j
ypredict.append(yl)
plt.scatter(x, y)
plt.plot(xpredict, ypredict)
plt.show()